python import numpy as np from sklearn.datasets import load_breast_cancer from sklearn.model_selection import train_test_split from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier from sklearn.svm import SVC from sklearn.metrics import accuracy_score from scipy.stats import entropy

加载数据

data = load_breast_cancer() X, y = data.data, data.target

切分：保留少量初始标签，大量作为未标注池

X_train, X_unlabeled, y_train, y_unlabeled = train_test_split( X, y, test_size=0.95, random_state=42, stratify=y )

模拟真实场景：未标注池的标签我们暂时不用（评估时除外）

print(f"初始标注样本数: {len(y_train)}, 未标注池大小: {len(X_unlabeled)}")

### 4.2 构建异质委员会
委员会成员必须保持多样性。这里我们选择随机森林、梯度提升树和 SVM 作为三个成员。为了让 SVM 输出概率，我们将 `probability=True`。

```python
def create_committee():
    return [
        ("RF", RandomForestClassifier(n_estimators=50, random_state=0)),
        ("GB", GradientBoostingClassifier(n_estimators=50, random_state=0)),
        ("SVM", SVC(kernel="rbf", probability=True, random_state=0))
    ]

4.3 计算投票熵分歧

实现一个基于投票熵的查询函数。对未标注池中的每一个样本，委员会成员分别预测类别，然后计算投票分布的熵。

def vote_entropy_query(committee, X_pool, n_instances=10):
    # 收集所有成员的类别预测 (n_committee, n_samples)
    votes = np.array([estimator.predict(X_pool) for _, estimator in committee])
    # 转置为 (n_samples, n_committee)
    votes = votes.T

    entropies = []
    for sample_votes in votes:
        # 计算每种票的计数
        _, counts = np.unique(sample_votes, return_counts=True)
        # 投票比例分布
        prob = counts / len(committee)
        ent = entropy(prob, base=2)   # 以2为底，也可用自然对数
        entropies.append(ent)

    # 选取熵最大的前 n_instances 个样本的索引
    query_idx = np.argsort(entropies)[-n_instances:]
    return query_idx

4.4 主动学习循环

我们进行多轮迭代，每轮选择 10 个争议样本添加到训练集，观察测试集上的准确率变化（此处用固定的留出集做评估，实际项目中建议使用验证集）。

# 留出固定的测试集用于最终评估
X_train_static, X_test, y_train_static, y_test = train_test_split(
    X, y, test_size=0.2, random_state=7
)

# 初始委员会训练
committee = create_committee()
for _, est in committee:
    est.fit(X_train_static, y_train_static)

# 模拟未标注池（实际中我们假装没有标签）
X_pool = X_unlabeled.copy()
y_pool_true = y_unlabeled.copy()  # 仅用于模拟标注

acc_history = []
n_queries = 20
batch_size = 10

for round_idx in range(n_queries):
    # 查询10个最高分歧样本
    query_indices = vote_entropy_query(committee, X_pool, batch_size)

    # 获取这些样本的“真相”（模拟人工标注过程）
    X_new = X_pool[query_indices]
    y_new = y_pool_true[query_indices]

    # 添加至训练集
    X_train_static = np.vstack([X_train_static, X_new])
    y_train_static = np.concatenate([y_train_static, y_new])

    # 从未标注池中移除已选样本
    X_pool = np.delete(X_pool, query_indices, axis=0)
    y_pool_true = np.delete(y_pool_true, query_indices, axis=0)

    # 重新训练委员会所有成员
    for _, est in committee:
        est.fit(X_train_static, y_train_static)

    # 评估：用简单投票法集成的准确率
    preds = np.array([est.predict(X_test) for _, est in committee])
    ensemble_pred = np.apply_along_axis(
        lambda x: np.bincount(x).argmax(), axis=0, arr=preds
    )
    acc = accuracy_score(y_test, ensemble_pred)
    acc_history.append(acc)
    print(f"第 {round_idx+1} 轮查询后 测试准确率: {acc:.4f}")